I've been writing about AI since before most people knew what it stood for, and still, barely a week goes by where I don't have to look up some new term or technical topic.
AI buzzwords are everywhere, but knowing what they actually mean can help you use AI better. So I put together an AI glossary that contains almost 100 of the most common, important, and (crucially) misunderstood AI terms.Â
The AI terms you should actually know
I've done my best to make these descriptions as easy to understand as possible without getting too into the weeds. But AI is a complex topic, and there's way more to it than what I've included here. If you want to dig deeper into anything, I've included lots of links, so click away.
A2A
A2A is an open agent-to-agent protocol (pioneered by Google) that enables independent AI agents to exchange messages and delegate tasks so they can collaborate seamlessly. A2A lets autonomous AI agents discover one another, advertise their capabilities, and exchange structured messages so they can work together in multi-agent systems without bespoke integrations. For a simple example, the idea is that your AI assistant agent would be able to book flights using a travel comparison website's AI agent.
Agentic AI
Agentic AI refers to autonomous AI systems that can set their own goals, plan, and execute sequences of actions—often coordinating tools or other agents—to achieve real-world objectives. Agentic AI systems use an observe → plan → act → observe loop to handle open-ended objectives. You could tell an agentic AI system to "secure me the best flight" or "debug this codebase," and it could do it with minimal further human prompting. Zapier Agents can help you build agentic AI systems that work across your entire tech stack.
AGI
Artificial general intelligence is a (still hypothetical) AI with human-level, domain-general understanding and the ability to learn and apply knowledge to any intellectual task. The threshold for what counts as AGI is fiercely debated, but no one yet claims it's been achieved. It's one of the few AI-related things that's still in the realm of science fiction, though I imagine we'll get closer and closer very quickly.
AI
Artificial intelligence is the broad field of building machines or software that perform functions we normally associate with human intelligence, from pattern recognition and language understanding to planning and decision-making. It spans symbolic logic, probability models, neural networks, evolutionary algorithms, and hybrid approaches, applied in domains such as chatbots, search, recommendation, medical diagnosis, and autonomous driving. Every other term on this list falls under the umbrella of AI, so it's a pretty fuzzy category to define.
AI agent
An AI agent is an AI system designed to autonomously perceive its environment, make decisions, and take actions toward achieving specific goals. AI agents can perform multi-step reasoning, remember context, and adapt to new situations with minimal human input, connecting with external tools via MCP and APIs. You can build your own AI agents using tools like Zapier Agents.
AI automation (intelligent automation)
AI automation (sometimes referred to as intelligent automation) combines automation technologies with AI to tackle tasks that require some level of intelligence. With Zapier, for example, you can set up an automation that triggers when a new customer support ticket is submitted; AI can analyze it for sentiment, and if it's negative, the automation will immediately flag it for you in Slack.
AI chatbot
An AI chatbot is a conversational system that uses AI language models to interact with users via text or speech. Modern chatbots, like ChatGPT, Google Gemini, and Claude, rely on large language models (LLMs) and can accomplish almost any task you could think of. You can also build your own AI chatbot using a tool like Zapier Chatbots and train it to work exactly how you want it to.
AI integration
AI integration embeds AI capabilities into an existing product, workflow, or infrastructure. Tools like Zapier make it possible to add AI to almost any service or workflow, so you can, for example, pull the power of ChatGPT into any of the other apps you use at work.
AI model
An AI model is a trained function that maps inputs to outputs based on patterns it's learned from data and training. AI companies release new models all the time, consistently improving on previous models. For example, within a matter of years, we saw GPT-3, GPT-3.5, GPT-4, GPT-4.5, o1, o3, o4, and dozens of other specialized model versions released from OpenAI alone.Â
AI model familyÂ
An AI model family is a set of related models that share an architecture and training recipe but differ in size or specialization. For example, while GPT is a model family, GPT-4o is a model. AI companies often have multiple model families. For example, OpenAI has both the GPT family and the o-series family.
AI orchestration
While AI automation helps complete individual tasks, AI orchestration manages the full workflow and links tasks into a cohesive system. It doesn't just notify sales when a lead comes in; it qualifies that lead, enriches the data, routes it to the right rep, and personalizes the next steps. It adapts as work unfolds, scaling impact across teams. Zapier is the leading AI orchestration platform, integrating AI and connecting across 8,000+ apps.
Alignment
Alignment is the effort to make an AI system's goals and behavior match human values and intent. It's what keeps new models from creating offensive, racist, and other objectionable content. To create alignment, AI developers use techniques like reward design, oversight, and constitutional training. As systems gain autonomy and open-ended skills, alignment will become more and more important.
ANI
Artificial narrow intelligence (ANI) excels at one task or domain but lacks general reasoning. Examples include spam filters, chess engines, and vision models that label animals. Each one outperforms humans in its niche but fails outside that niche. Nearly all commercial AI today is considered ANI.
Attention
Attention is a neural-network mechanism that lets a model weigh which input elements matter most when producing each output. First used in translation, it evolved into the transformer's self-attention, which is at the core of modern large language models (LLMs). It's a tricky concept to grasp, but the big takeaway is that it's one of the key developments that led to the latest AI models.
AutonomousÂ
Autonomous describes a system that can perceive, decide, and act in its environment without human oversight. Autonomy covers everything from a Roomba mapping a room to a trading bot managing a portfolio to ChatGPT doing a web search if it doesn't know the answer to something. It requires sensing, planning, and execution modules, and it (hopefully) has safeguards and override channels in case things go awry.
Bias (variance)
Bias is the systematic skew, omission, or mislabeling in a dataset that causes an AI model to learn a distorted view of the real world. It creeps in when certain classes, demographics, regions, time periods, or edge cases are over- or under-represented, or when labels reflect human prejudices and recording errors. For example, only training AI models on data from one country or one community can make it act as if that country or community is the only one that exists.
Benchmark
A benchmark is a standardized test set and metric used to compare AI models on a task. Examples include MMLU Pro, GPQA Diamond, Humanity's Last Exam (core knowledge and reasoning), MMMU (multimodal understanding), and HumanEval (code generation). This is how technical folks compare which models are the best at which tasks—and how we know if the latest OpenAI model can ace the LSAT.
Chain-of-thought reasoning
Chain-of-thought reasoning makes language models generate step-by-step explanations before arriving at the answer. Using intermediate logic guides the model toward better solutions to harder problems. It can be prompted (e.g., "Let's think step by step") or learned via supervision as it is in reasoning models.
ChatGPT
ChatGPT is OpenAI's conversational chatbot. It launched publicly in November 2022 and popularized AI chatbots, giving the general public an easy way to experience the power of advanced AI models. It's generally at the forefront of AI development, consistently adding new models and functionalities.
Claude
Claude is two things: (1) Anthropic's large language model family designed for helpful, harmless, and honest conversation and (2) the AI chatbot built on that model family. It's known for its advanced coding abilities and its early ability to create apps on the chatbot.
Compute
Compute is the total processing power used to train or run AI models (it's measured in FLOPS or GPU hours). Generally, we relate compute to model quality—as one increases, so does the other—but the relationship is more complicated. Compute is the most expensive part of training models because running the computer hardware requires a lot of electricity.
Computer vision
Computer vision lets machines and AI applications interpret and act on visual data such as images and video. It's what lets ChatGPT look at an image and understand what's in it, for example, but it can also be used for things like medical imaging and autonomous driving.
Computer useÂ
Computer use is a feature of some AI applications, where they can understand typical operating system and app user interfaces and interact with them by simulating input devices like keyboards, mice, and game controllers. This means, for example, that an AI chatbot could open a file on your computer for you.
Context length (context window)
Context window is how much history or text an AI model can process at once. More technically, it's the maximum number of tokens an LLM can consider in one request. Context length is a synonym for context window, but it also refers to the current number of tokens currently in use. For example, a prompt could have a context length of 3,000 tokens even if the model has a context window of 128,000 tokens. The larger the context window, the more context a model can incorporate as it creates its outputs. For example, you could upload an entire novel to some AI models with long context windows, and they'd be able to consume all of it in one go. A smaller context window couldn't do that.
Decision-making mechanisms
Decision-making mechanisms are the algorithms that choose what action an AI system will take from all the options available to it. In AI, they include Bayesian reasoning, reinforcement-learning policies, and planning search trees. Good decision-making mechanisms are able to weigh up the pros and cons of different approaches and act in transparent and fair ways. For example, self-driving cars use decision-making mechanisms to decide when to change lanes, slam on the brakes, or proceed at the speed limit.
Deep learning
Deep learning is part of machine learning—a "deep" part, in that the computers can do even more autonomously, with less help from humans. The massive dataset that the computer is trained on is used to form a deep learning neural network: a complex, many-layered, weighted algorithm modeled after the human brain. That means deep learning algorithms can process information (and more types of data) in an incredibly advanced, human-like way. While ChatGPT is an incredible tool, even its developers don't fully understand what happens inside the system.
DeepSeek
DeepSeek is both a series of open AI models released in 2024 and the Chinese AI company that developed them. Models in the family include DeepSeek V3 and the DeepSeek R1 reasoning model. Unlike most popular models, it's open—which means anyone can download the model and run it on their own hardware. DeepSeek caused a kerfuffle when it was first released because it became clear that advanced AI models could be developed by non-American companies, at a fraction of the cost.
Dense modelÂ
A dense AI model activates all of its parameters for every input (unlike sparse or mixture-of-experts [MoE] models). Dense models are simpler to train, but they require more compute the bigger they get.
DistillationÂ
Distillation trains a small "student" model to mimic a larger "teacher" model's outputs. Many smaller models are distilled from larger models rather than trained independently. For example, Llama 4 Maverick (400B total parameters) and Llama 4 Scout (109B total parameters) were distilled from Llama 4 Behemoth (2T total parameters).
Edge device
An edge device is a resource-constrained computer, smartphone, drone, or any other device running AI near the data source. Edge inference cuts latency, saves bandwidth, and preserves privacy. It requires smaller models and, often, dedicated computer chips.
Embedding (vector embedding)
An embedding (or vector embedding) encodes semantic information about data. To simplify it, vector embeddings are a way of turning data, like the pages on your website, into numbers so a computer can understand how they're related. The point of vectorizing data is to make it searchable by meaning across a bunch of different dimensions. Similar items map to nearby points, enabling search and clustering. The idea is that things like "king – man + woman ≈ queen." LLMs produce text embeddings for retrieval-augmented generation (RAG) or recommendation.Â
Few-shot prompt
Few-shot prompting is an example-based prompting technique—it sits between zero-shot and many-shot approaches by giving the model just enough examples (typically 2–5) to anchor format, style, or reasoning without exhausting the context window. Because it's done entirely at inference time, few-shot prompting is a fast, code-free way to adapt large language models to new tasks.
Fine-tuning
Fine-tuning is kind of like training an AI model for your specific end goal. When you fine-tune a model, you adjust a pretrained model's weights using a smaller task-specific dataset. For example, chatbots like ChatGPT use models that are fine-tuned on examples of conversation. Fine-tuning requires far less compute than training from scratch.Â
GPTÂ
GPT (generative pretrained transformer) is OpenAI's model family pretrained on vast amounts of text (and, eventually, other data) and fine-tuned for tasks like chat and coding. GPT-1 was launched all the way back in 2018, but it was nothing like the GPT models we have access to now. GPT models power ChatGPT and many other AI tools.Â
GPUÂ
A GPU (graphics processing unit) is a type of parallel processor that can train and run neural networks much faster than a CPU. Its ability to perform many calculations simultaneously makes it ideal for processing the large volumes of data required by AI models. As a result, GPUs are widely used to power tools like AI chatbots, image recognition systems, and other machine learning applications.
Guardrails
Guardrails are safety filters and policies that block or modify unsafe or undesired AI model outputs. They detect things like profanity, disallowed content, or jailbreak attempts. Guardrails can run before, during, or after generation. As you'll see any time an AI chatbot goes awry, balancing safety with over-blocking remains challenging.
Gemini
Gemini is Google's multimodal AI model family—and also Google's AI chatbot that runs on the model family. Gemini models include Nano, Flash, Pro, and Ultra, and each is designed for a different balance between price, performance, and speed.
Generative AI
Generative AI creates new content, like text, images, audio, or code, rather than merely analyzing data. It's used in everything from AI chatbots to image generators to really any app that can produce text. Of course, with generative AI, ethical debates are everywhere—including concerns about originality, bias, misuse, and how training data was acquired.
Google AI Mode
Google AI Mode is built into Search and adds a conversational, chatbot-like experience to your Google searches. It lets you ask complex, multi-part questions and receive AI-generated responses scraped from the web. And like a chatbot, the feature also supports follow-up questions, so your search experience feels more interactive and tailored to you.
GrokÂ
Grok is xAI's conversational (and controversial) LLM. While its marketing emphasizes its humor and minimal censorship, its real world performance is more or less in line with other top models. It also integrates directly with X.
Hallucination
A hallucination is a confident but false statement generated by an AI model; for example, the idea that glue is a useful pizza ingredient. Hallucinations occur when the model completes a pattern without verifying its accuracy. All generative AI models are susceptible to some level of hallucinations, but they've become less of an issue the more advanced the models become. They can be mitigated by retrieval augmented generation, tool use, and additional verification steps.
Inference
Inference is the phase where a trained model takes new inputs to produce outputs. Unlike training, which iteratively updates weights and is done offline, inference happens in the production deployment stage. Whenever you use ChatGPT, it's running inference.
InterpretabilityÂ
Interpretability is the degree to which humans can understand why an AI system produced a given output. If you can't tell why an AI model has responded in a particular way, it's hard to fix it or ensure its reliability in critical systems. Interpretability methods—which increase trust, debugging, bias detection, and regulatory compliance—include things like saliency maps, feature importance, and examining distilled models.
Jailbreak
A jailbreak is a prompt or exploit that bypasses an AI system's safety filters. Tactics range from getting a chatbot to role-play as a criminal to Unicode tricks. When there are known jailbreaks, AI providers can patch them and add detectors, but it's really an ongoing cat-and-mouse game.
Latency
Latency is the time between sending an input to a model and receiving the output—when ChatGPT takes a few moments to respond, that's latency. How long it takes depends on a lot of factors, including model size, hardware, and network speed. There are ways to cut latency, like edge deployment, but generally it's not too bad as an end user (given what's actually happening behind the scenes).Â
Learning systems
Learning systems improve AI performance over time by updating models or policies based on data or feedback. They contrast with rule-based programs, which remain static.
LLM
A large language model (LLM) is an AI model trained on massive amounts of text so it can predict the most appropriate next token (output). LLMs can be prompted, fine-tuned, or used as tool-calling agents, and they power many of the most popular AI tools you'll encounter, like chatbots and AI assistants. For a better understanding of how LLMs work, learn more about how ChatGPT works.
Llama
Llama is Meta's open-weight LLM family. It's basically the Facebook parent company's response to OpenAI and Google Gemini, but with one key difference: all the Llama models are freely available for almost anyone to use for research and commercial purposes. It's credited with kickstarting interest in open AI models that you can download to your own server and train however you want.
Local
Running an AI model locally means executing it entirely on a user device or private server. Local deployment offers privacy, offline access, and no inference fees. The main limitation is hardware capacity, though some small models can run effectively on consumer hardware.
Machine learning (ML)
Machine learning is a subfield of artificial intelligence. Instead of computer scientists having to explicitly program an app to do something, they develop algorithms that let it analyze massive datasets, learn from that data, and then make decisions based on it. Machine learning methods include supervised, unsupervised, and reinforcement learning, and ML forms the basis of many of the recent developments in AI, including transformer models.
MCP
MCP is an open, client-server standard that lets AI applications connect to external data and tools. For example, it allows you to query your CRM database without leaving Claude. By using a single protocol, developers can avoid having to build custom integrations for each app. Pioneered by Anthropic, MCP has been embraced by Google, OpenAI, and other leading AI providers. Zapier MCP makes it even easier, letting you connect your AI tools to 8,000+ apps.
Mixture of experts (MoE)
A mixture-of-experts model is any machine learning model composed of multiple smaller specialized models (experts) and a gating or routing network to select which expert is used for any given input. An easy way to think of it is that a regular (or dense) AI model has one super intelligent expert, while a MoE model is a team with multiple more specialized experts, plus a manager who decides which experts solve which problems. For example, an MoE model with 400B total parameters may only activate 17B at any one time. While more complicated to develop than dense models, they offer clear performance benefits.Â
Modality
Modality refers to a type of data, like text, vision, or audio. Models can have different input and output modalities; for example, they might be able to read images but only output text in response. Multimodal models are models that can process multiple data sources.
Multimodal AI
Multimodal AI processes and links multiple types of data (e.g., text, vision, audio, code) to reason or generate across them. For example, multimodal models might be able to describe an image using text or audio or create images from text prompts. At this point, a lot of the top AI models are multimodal.
Natural language processing (NLP)
NLP is the process through which AI is taught to understand the rules and syntax of language, programmed to develop complex algorithms to represent those rules, and then made to use those algorithms to carry out specific tasks, like language generation, translation, and summarization.
Neural network
A neural network is a kind of computer algorithm modeled off the human brain, and it's typically created using machine learning or deep learning. Neural networks can approximate complex functions without being programmed directly, and they're at the core of most modern AI models. Neural networks need to be trained on vast quantities of data to work effectively.Â
o-series
o-series is Open AI's family of reasoning models (including o1, o3, and o4, along with many variations). They use additional reasoning tokens to generate a chain of thought that allows them to better solve complex problems.
One-shot prompt
One-shot prompting is a type of example-based prompting (right between zero-shot and few-shot), where the user provides a single example to the model to help it understand the desired task or format. Performance naturally hinges on choosing a solid representative case.
Open source AI
Open source AI releases model code, weights, or data under permissive licenses for anyone to use and study. Depending on the specifics of the license used and what's released, AI models can be referred to as open source, open weight, or open. Most popular models are proprietary, but Meta's Llama family and DeepSeek are notably open.
OpenAI
OpenAI is the AI research and deployment company behind ChatGPT. Along with Google, Meta, Anthropic, and DeepSeek, it's one of the major AI providers. OpenAI CEO Sam Altman is all over AI news and has been one of the most important and polarizing figures in AI development.
Parallelization
Parallelization splits model training or inference across multiple processors to finish faster. Instead of using a single GPU or other chip, models like GPT-4o and Llama 3 can run on a stack of GPUs for better performance.
Parameters
Parameters are variables in a model—adjustable weights that capture the knowledge the AI model has learned about the world. Training, fine-tuning, and alignment determine the value of the parameters; more parameters capture more complexity but also require more compute. The quality of an AI model isn't purely proportional to parameter count, but it is a major factor. LLMs typically have billions of parameters. Â
Perplexity
Perplexity is an AI-powered search engine that lets you ask questions in natural language and replies with a short answer plus source links. It mixes a traditional web index with large language models to read, rank, and summarize results on the fly. Each answer comes with clickable citations so you can verify where the facts came from. Google's AI Mode does something similar now, but Perplexity was purpose built as an AI search engine.
Pretraining
Pretraining is the large-scale learning phase that gives an AI model broad knowledge before fine-tuning. LLMs are pretrained to predict the next token using a massive dataset that includes the whole public internet and any private, licensed, and synthetic data the developers can find or create. Pretraining is a major portion of the compute cost for LLMs.
Processors
Processors are hardware units that execute the math behind AI. The main ones for modern AI models are GPUs and TPUs.
Prompt
A prompt is the input or instructions supplied to a language model to elicit a response. It's what you say to a chatbot to get it to respond.
Prompt engineeringÂ
Prompt engineering is the art (or science, depending on how you look at it) of designing prompts to steer an AI model toward the desired behavior. It includes even simple things like offering role instructions or giving examples to try to get the best response. Despite model advances, prompt engineering remains a key lever you can use when working with AI.
Prompt injection
Prompt injection is an attack where adversarial text overrides the system prompt to make the model misbehave, exploiting a model's tendency to obey the latest instruction. Methods include splitting a malicious prompt into multiple components that each appear innocent, adding malicious instructions to uploaded files or images, or injecting malicious instructions through external tools.
Proprietary models
Proprietary models are closed-weight AI systems whose parameters and data aren't publicly released. Instead, the companies behind these models expose them via APIs under license terms, which means you can use the models, but you can't really build off them the way you could with an open model. OpenAI, Anthropic, and Google primarily develop proprietary models.Â
Qwen
Qwen is Alibaba's open LLM family, and it includes text, code, and vision variants. Qwen3, the latest version, is one of the largest open mixture-of-experts models.Â
Red teaming
Red teaming is a practice where humans actively probe an AI system for failures or unsafe behavior by crafting adversarial prompts and scenarios. In other words, the developers try to break their own models before bad actors get the chance to.
Reinforcement learning
Reinforcement learning is a process where a model learns to be more accurate by having its outputs assessed, either by a human or by another AI designed to rank them. Reinforcement learning from human feedback (RLHF) applies the same idea to align LLMs with human preferences. It's essentially a reward model with comparison data (where two or more model responses are ranked by AI trainers) so the AI can learn which was the best response in any given situation. It's an important training tool that helps AI models get better and fine-tuned effectively.
Responsible AI
Responsible AI aims to build and deploy AI systems that are ethical, fair, transparent, and accountable. Frameworks address bias testing, privacy, governance, and environmental impact, and standards such as the EU AI Act codify best practices.
Retrieval augmented generation (RAG)
Retrieval augmented generation or RAG couples a language model with a retrieval step (powered by vector databases and embeddings) so that AI generation is grounded in external documents. For example, a customer service chatbot might pull content from the employee handbook. At query time, the system fetches relevant passages and injects them into the prompt. This boosts factual accuracy and allows updating knowledge without retraining.Â
Semi-supervised learning
Semi-supervised learning (SSL) trains a model on a mix of labeled and unlabeled data. It reduces annotation requirements while improving performance over purely supervised data. SSL is common in vision and speech models.
Small language models
Small language models are just small large language models. I know that sounds kind of silly, but as large language models have become larger and (as a result) more powerful, there's been the need for a handy term to categorize small, lightweight language models that still use the same state-of-the-art technologies. (One way to measure the size of language models is with the number of parameters they have.)
Stable Diffusion
Stable Diffusion is a family of open image generation models. Stability AI, the company behind Stable Diffusion, is now pushing its own ChatGPT alternative called Stable Assistant, but you can access earlier versions of Stable Diffusion through most AI art generators and lots of other tools that have an integrated image generator. You can also license the latest version of Stable Diffusion, install it on your own server, and even train it on your own data.
Strong AI
Strong AI (similar to AGI) is AI that would match or exceed human general intelligence and consciousness. It could learn, reason, and adapt across any domain. Good news: no existing system is strong AI.
Supervised learning
Supervised learning trains a model on input-output pairs so it can predict the best outputs for new inputs. It uses structured, labeled datasets, and the quality depends on clean, representative labels. While supervised learning can be effective in some circumstances, the training datasets are incredibly expensive to produce. Even now, there just isn't that much data suitably labeled and categorized to be used to train LLMs.
Synthetic data
Synthetic data is artificially-generated information used to train or test models when real data is scarce or sensitive.
TokenÂ
A token is the unit that a language model processes. While all data, including images and audio, are also broken down into tokens, the concept is simplest to grasp with text, where a token is typically a word fragment (though it can be whole words, punctuation marks, and other things too). Tokenization splits text into tokens before embedding, and models generate outputs one token at a time. GPT-3, the original model behind ChatGPT, was trained on roughly 500 billion tokens, which allowed its language models to more easily assign meaning and predict plausible follow-on text by mapping them in vector-space.
TokenizationÂ
Tokenization is the process of converting raw text into a sequence of tokens for model input.
TPU
A TPU (tensor processing unit) is Google's custom chip optimized for AI workloads. It's designed specifically to speed up machine learning tasks.
Training data
Training data is the information a model learns from during training. In a lot of cases, this includes the entire open internet, plus any other proprietary resources the developers can get their hands on. The size, quality, and bias of training data will heavily influence the performance of the AI model.
Transformer architecture
Transformer architecture is a type of deep learning model that reads all words in a sentence at once and uses a mechanism called self-attention to identify relationships between them, no matter their position. Unlike older models that process text sequentially, transformers perform computations in parallel, making training faster and more efficient. The architecture, introduced in 2017, is foundational to modern AI models like ChatGPT and has made them both more powerful and cost-effective to develop.
Unsupervised learning
Unsupervised learning finds patterns in unlabeled data using techniques like clustering, dimensionality reduction, and generative modeling. What this means is that the AI crunches through a massive amount of data to develop its own understanding of the rules and relationships that govern that data.
Vector encoding (vector embedding)
A vector encoding (often called an embedding) is a dense list of numbers that an AI model produces to capture the meaning or visual characteristics of an item—the end goal is that similar items will end up with nearby numbers in the same space. Because the vector math preserves relationships ("king – man + woman ≈ queen"), these encodings power tasks like search, clustering, recommendation, and few-shot learning.
Vector database
A vector database is a specialized datastore that indexes and retrieves high-dimensional vectors using similarity search (nearest-neighbor) algorithms instead of traditional key or SQL look-ups. They're the backbone of retrieval-augmented generation (RAG), semantic search, recommendation engines, and real-time analytics, where "find items like this one" need to happen instantly.
Weak AI (narrow AI)
Weak AI or narrow AI performs specific tasks without generalized understanding—for example, spam filters and image classifiers. They can surpass humans in their slice, but they fail outside it.
xAI
xAI is Elon Musk's AI company that developed the AI model family Grok. It now owns X, the social network formerly known as Twitter.
Zero-shot prompting
Zero-shot prompting is a prompting technique where you ask a model to perform a task without giving it any examples of how to do it. Instead, the model relies only on the instructions or phrasing of the prompt, along with its prior training, to understand and complete the task. It's pretty common when you're chatting with an AI chatbot.
Related reading:
